ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels
- PDF / 2,514,599 Bytes
- 42 Pages / 439.37 x 666.142 pts Page_size
- 95 Downloads / 186 Views
ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels Angus Dempster1
· François Petitjean1
· Geoffrey I. Webb1
Received: 28 October 2019 / Accepted: 18 June 2020 © The Author(s), under exclusive licence to Springer Science+Business Media LLC, part of Springer Nature 2020
Abstract Most methods for time series classification that attain state-of-the-art accuracy have high computational complexity, requiring significant training time even for smaller datasets, and are intractable for larger datasets. Additionally, many existing methods focus on a single type of feature such as shape or frequency. Building on the recent success of convolutional neural networks for time series classification, we show that simple linear classifiers using random convolutional kernels achieve state-of-the-art accuracy with a fraction of the computational expense of existing methods. Using this method, it is possible to train and test a classifier on all 85 ‘bake off’ datasets in the UCR archive in < 2 h, and it is possible to train a classifier on a large dataset of more than one million time series in approximately 1 h. Keywords Scalable · Time series classification · Random · Convolution
1 Introduction Most methods for time series classification that attain state-of-the-art accuracy have high computational complexity, requiring significant training time even for smaller datasets, and simply do not scale to large datasets. This has motivated the development of more scalable methods such as Proximity Forest (Lucas et al. 2019), TS-CHIEF
Responsible editor: Aristides Gionis, Carlotta Domeniconi, Eyke Hüllermeier, Ira Assent.
B
Angus Dempster [email protected] François Petitjean [email protected] Geoffrey I. Webb [email protected]
1
Faculty of Information Technology, Monash University, Melbourne, Australia
123
A. Dempster et al.
Fig. 1 Mean rank of Rocket versus state-of-the-art classifiers on the 85 ‘bake off’ datasets
(Shifaz et al. 2020), InceptionTime (Ismail Fawaz et al. 2019c), MrSEQL (Le Nguyen et al. 2019), and cBOSS (Middlehurst et al. 2019). We show that state-of-the-art classification accuracy can be achieved using a fraction of the time required by even these recent, more scalable methods, by transforming time series using random convolutional kernels, and using the transformed features to train a linear classifier. We call this method Rocket (for RandOm Convolutional KErnel Transform). Existing methods for time series classification typically focus on a single representation such as shape, frequency, or variance. Convolutional kernels constitute a single mechanism which can capture many of the features which have each previously required their own specialised techniques, and have been shown to be effective in convolutional neural networks for time series classification such as ResNet (Wang et al. 2017; Ismail Fawaz et al. 2019a), and InceptionTime. In contrast to learned convolutional kernels as used in typical convolutional neural networks,
Data Loading...
